知の物理学研究センター / Institute for Physics of Intelligence (iπ)

【日時/Date】
2023年12月27日（水）10時30分～12時 / Dec. 27, 2023, 10:30 - 12:00 (JST)

【発表者/Speaker】
深井 朋樹 氏（沖縄科学技術大学院大学）

【会場】
理学部１号館９階９１３号室

【タイトル/Title】
“Graphical information representations in associative memory models”

【概要/Abstract】

The associative memory model attempts to explain the brain's memory processing using attractors, and various studies have been conducted from the viewpoints of physics, engineering, psychology, biology, etc. This classic model has not long been given much attention, but it has recently been revived as its relationship with large-scale language models has been pointed out. After a brief introduction to classical and recent results of associative memory models, I will present two recent studies from my group in this seminar. First, I will show correlated-attractor associative memory models embedding graphical relationships between memory states. In the brain, memories are rarely mutually independent but usually have some hierarchical relationships. Our associative memory model stores hierarchical relationships between memory states using correlated attrator states. In terms of machine learning, our model operates similarly to the graph Laplacian for analyzing the hierarchical cluster structure of graphs. Then, the simplicial Hopfield model attempts to improve memory storage ability by mixing higher-order interactions (two-body, three-body, etc.) between neurons, such as represented by the simplicial complex. Associative memory models with many-body interactions have been studied, but their superiority in memory recall has not been demonstrated on large-scale databases such as Image Net. Finally, the relationship has been recently pointed out between spatial navigation by the hippocampus and conceptual processing and natural language processing. If time permits, I will also touch upon the results of such studies.